existAI for ScienceIn this field, we have been thinking: can scientific objects such as proteins, small molecules, materials and chemical reactions be modeled and generated by a unified model, just like human language?
The current mainstream approach to AI4S is to train a separate expert model for each problem. There are specialized models for structure prediction, molecular generation, and retrosynthetic programming. Each performs its own function.Cross-domain knowledge is difficult to transfer and reuse.
More importantly, they are mostly built on a setindependentIt is based on the technology stack of large language models: relying on scarce 3D coordinate data and dedicated geometric networks, and is completely separate from the mature engineering infrastructure already existing in the LLM ecosystem.
But LLM has already proven its effectiveness in the field of natural language processing."One model solves thousands of tasks"The feasibility of this is under consideration. Could AI for Science also follow this path?
LOGOS(Language Of Generative Objects in Science) provides a specific path: design a "scientific grammar" that allows multi-domain scientific data to run natively within a large language model framework.
What is LOGOS?
It was open-sourced by ATH-Token Foundry in collaboration with the Hillhouse Institute of Artificial Intelligence at Renmin University of China.A multi-domain science generation foundation model based on a unified "scientific grammar".
The core insight isProteins, small molecules, and materials may appear different on the surface, but they fundamentally follow similar rules of composition, structural constraints, and interaction semantics. They are essentially different dialects of the same "scientific language." The value of a unified grammar lies precisely in this: cross-domain knowledge transfer, multi-task collaborative optimization, and alignment of pre-training with downstream goals all become possible because of a common linguistic foundation.
Performance Overview: Six Tasks Explained in One Image
Before delving into the technical details, let's first look at LOGOS's "real-world performance." In six representative scientific tasks, LOGOS, using a pure sequence modeling paradigm, consistently matched or surpassed domain-specific methods.
Of particular note is itsParameter efficiencyLOGOS-1B achieves high performance across multiple tasks with only 1/56th the number of parameters (1B vs 8×7B).Surpassing NatureLMIt demonstrates extremely high parameter efficiency.
Experiment Highlights: A Step-by-Step Explanation of the Major Tasks
🟣 Task 1: Pocket Conditional Ligand Generation — Pure Sequence Model Beats 3D Diffusion Model for the First Time
This is the core application scenario of AI-driven drug development: given a protein pocket, it generates small molecule drugs that can bind precisely.
LOGOS-1B uses only about1/56Its parameter count surpasses that of NatureLM (8×7B), demonstrating extremely high parameter efficiency.
In the pure sequence paradigm, it outperformed diffusion models that rely on 3D coordinates, demonstrating the effectiveness of transforming spatial structures into sequences. The resulting molecules simultaneously meet the triple requirements of binding affinity, drug similarity, and synthetic feasibility, exhibiting potential for practical drug design.
🟣 Task 2: Retrosynthetic Prediction – Directly Assisting Chemical Synthesis Route Planning
Given a target molecule, predicting the precursor molecules required to synthesize it is a core tool for chemists when designing synthetic pathways.
LOGOS achieved a Top-1 accuracy of 100%.74.8%,Completely surpassAll baseline methods. This means chemists can directly use LOGOS to assist in designing synthetic routes, reducing trial-and-error costs.
🟣 Task 3: Pocket Site Identification — Predicting Protein Binding Sites Based on Sequence Only
Predicting where small molecules can bind to a protein's surface is the starting point for drug discovery. Traditional methods rely on 3D structural data, but many proteins simply do not have usable 3D structures.
LOGOS achieves high performance on the HOLO4K dataset using only amino acid sequences.58.5%Its Top-n accuracy surpasses all traditional 3D methods (except P2Rank). This result extends pocket recognition from a "3D structure-dependent" paradigm to a new paradigm that "only requires sequence information".
🟣 Task 4: MOF Material Generation – Exploring the Unknown Chemical Space
The development of novel metal-organic framework materials has important applications in gas storage, separation, catalysis and other fields.
The NBB (new building block ratio) generated by LOGOS increased from 10.10% in MOFFlow-2 to 17.78%, an improvement of [percentage missing].76%The model possesses the ability to transcend existing components in the training data and generate entirely new molecular building blocks. AI's contribution to science should not stop at reproducing the known, but should extend to the generation and creation of the unknown.
The other two tasks: protein editing (Hard difficulty, Fitness score reached 0.93, compared to baseline 0.34).Increased by 174%The antibody CDR design (with an AAR of 79.82%, surpassing the structure-dependent backfolding method RADAb) also outperforms the strongest existing baseline, further validating LOGOS's generalization ability in a wider range of scenarios.
Where is the core breakthrough? Three key designs.
LOGOS selects modalities strictly according to the scientific community's "knowledge system".
Core trunk
Proteins are the core executors of life activities, while small molecules are the most important chemical means of regulating protein function and intervening in life processes. They are two core areas of scientific research.
Family expansion
Antibodies are special proteins with high sequence diversity and specific recognition capabilities, and are an indispensable member of the biological macromolecule family; chemical reactions and functional materials further expand the knowledge coverage of the small molecule chemical space from the perspective of transformation processes and material composition.
Interface Interaction
Proteins and small molecules do not exist in isolation. The interfacial interaction between them—especially the selective recognition of small molecule ligands by protein binding pockets—constitutes a key link connecting biological macromolecules and chemical entities, and is also the basis for understanding molecular recognition mechanisms and drug action mechanisms.
Based on this logic, LOGOS has built a comprehensive...7 types of modes,total44.87BA massive pre-trained corpus of tokens:
Biomolecular layerProtein (28.9 B tokens) + Antibody (3.0 B tokens)
Chemical entities and transformation layersSmall molecules (2.1B tokens) + chemical reactions and MOF materials (0.47B tokens)
Interface Interaction LayerProtein pocket (5.8 B tokens) + Protein pocket-ligand complex (4.6 B tokens)
This ensures that this "dictionary" not only has a large vocabulary, but also conforms to real scientific principles.
🟣 Innovation Point 1: Unifying "Scientific Grammar" + Discretization of Spatial Interaction
This is LOGOS' most groundbreaking design, consisting of two complementary parts.
Technology A: A unified vocabulary, everything can be a token.
LOGOS designed a shared vocabulary that encodes heterogeneous objects such as proteins, small molecules, and materials—which were originally incompatible—into a unified discrete token sequence. This allows them to be understood and generated by a large model in an autoregressive manner within the same generative space.
Technology B: Understanding Space Through Sequences Without 3D Coordinates
Traditional AI relies on explicit 3D coordinates and complex geometric neural networks to understand how proteins and small molecules bind. LOGOS has invented a "textual description method": it directly "syntacticizes" 3D spatial contact patterns into discrete tokens.
The three-dimensional relationship of the protein pockets was written as three progressive "textual sequences":Residue-pocket labeled sequences:First, label the basic information at the amino acid level;Residue-pocket-small molecule unfolded sequence: Incorporating surface contact information to make the structure more three-dimensional;Residue-pocket switching sequence:A special marker is used to force amino acids and molecular structures to be "aligned" in the sequence.
The result is that LOGOS does not need to input 3D coordinates at all; it can construct complex 3D spatial interaction rules in the mind simply by "reading text" (sequence prediction).
🟣 Innovation Point Two: Pre-training and downstream tasks have completely consistent format and objectives.
In traditional paradigms, changing a research stage (such as from structure prediction to molecular generation) often requires a new model and new assumptions. At the same time, there is a significant "target bias" between the pre-training objectives and the downstream tasks. This "disconnect between learning and application" leads to a large number of fine-tunings required when the model is deployed.
LOGOS's scientific syntax design completely solves this problem:
Consistent in form:The sequence form of the pre-training data is equal to the input-output form of the downstream task.
Consistent goals:The pre-trained next-token prediction is equal to the downstream conditional generation target.
This form-objective alignment effectively eliminates the gap between pre-training and downstream applications, activating generative capabilities without the need for complex adaptation layers or extensive fine-tuning.
🟣 Innovation Point 3: Cross-domain knowledge positive transfer, proven effective by experiments.
A unified syntax not only allows scientific subjects to "communicate in one language," but also enables them to "share knowledge" at the underlying level. To verify whether LOGOS has truly learned cross-domain knowledge, the team conducted in-depth research.
Validating "cross-language translation" capabilities
The model can see the "dialect" (amino acid pocket sequence) of proteins and directly "translate" it into the "dialect" (SMILES structure) of small molecules. This proves that it has truly learned the correspondence between the two.
Verify the necessity of uniform syntax components
The team conducted a "deconstruction experiment" to verify the value of each design step: if pocket and complex data were removed, the model performance was close to random (Vina -3.57); after adding all grammatical components, the performance jumped significantly (Vina -7.64).
Validating the synergistic effect of "1+1>2"
The model learns four tasks simultaneously, which is more effective than training them independently, demonstrating positive cross-domain knowledge transfer.
“
LOGOS's practical experience has provenThe development of AI for Science does not lie in rebuilding an independent technology stack separate from LLM (Language Modeling). Rather, it lies in deeply aligning fundamental scientific models with large language models.Shared architecture, shared training paradigm, shared inference infrastructure—This will make the basic model a true new entry point for AI4S.
"
By inheriting the pre-trained weights of LLM, scientific models are able toDirect reusevLLM's mature engineering infrastructure, such as accelerated inference and model quantization, significantly reduces training and deployment costs. Simultaneously, a unified syntax space...Breaking down disciplinary barriersThis allows data from different fields such as biology, chemistry, and materials to be fed back for training within the same model framework, forming a cross-domain knowledge loop.
From "selecting the known" to "designing the unknown," LOGOS is driving natural science research towards...Design-drivenandDiscovery-drivenTransformation. In the future, the team will further expand this "scientific grammar" to more fields such as genomics, and explore the deep integration of scientific language and natural language, striving to build a truly universal scientific foundation model.
Welcome to LOGOS. We have fully open-sourced the model weights, inference code, and technical report. You can obtain them as follows:
🤗 HuggingFace: https://huggingface.co/LOGOS-Hub
💻 GitHub: https://github.com/LOGOS-Hub/LOGOS
📄 Technical Report (Paper): https://arxiv.org/abs/2606.16905